Abstract
Specular highlights negatively affect photogrammetric 3D reconstructions. To mitigate this problem, we developed an AI-driven image processing technique able to remove specular highlights. We created a synthetic image dataset that reflects the objects, viewpoints, and specular behaviors found in real-world photogrammetric campaigns, and used it to train a U-Net model that can batch-process input images for photogrammetric reconstruction. The process was tested on both synthetic and real-world photos, demonstrating superior results compared to existing models in the literature.
